rayzee 9.0.0 → 9.1.0

package/src/Processor/BVHRefitter.js

@@ -1,3 +1,4 @@
+
 /**
  * BVHRefitter — Fast O(N) bottom-up BVH AABB refit for animated geometry.
  *
@@ -8,6 +9,8 @@
  * Designed to run in both main thread and Web Worker contexts.
  */
 
+import { isFoldedRef, foldedFirst, foldedCount } from './BVHLeafFold.js';
+
 // Inline copy of layout constants (source of truth: EngineDefaults.js).
 // Cannot import because this runs inside Web Workers where window is not defined.
 const TRIANGLE_DATA_LAYOUT = {
@@ -161,6 +164,98 @@ function transformBoundsToWorld( bvhData, nodeOff, src, srcOff, dst, dstOff ) {
 }
 
 
+/** Bounds of triangles [first, first + count) into out[ off .. off + 6 ). */
+function triangleBounds( acc, first, count, out, off ) {
+
+	const A = TRIANGLE_DATA_LAYOUT.POSITION_A_OFFSET;
+	const B = TRIANGLE_DATA_LAYOUT.POSITION_B_OFFSET;
+	const C = TRIANGLE_DATA_LAYOUT.POSITION_C_OFFSET;
+	let minX = Infinity, minY = Infinity, minZ = Infinity;
+	let maxX = - Infinity, maxY = - Infinity, maxZ = - Infinity;
+
+	for ( let t = 0; t < count; t ++ ) {
+
+		const gi = first + t;
+		const f = acc.chunked ? acc.f.chunkFor( gi ) : acc.f;
+		const o = acc.chunked ? acc.chunked.baseOf( gi ) : gi * FPT;
+		minX = Math.min( minX, f[ o + A ], f[ o + B ], f[ o + C ] );
+		minY = Math.min( minY, f[ o + A + 1 ], f[ o + B + 1 ], f[ o + C + 1 ] );
+		minZ = Math.min( minZ, f[ o + A + 2 ], f[ o + B + 2 ], f[ o + C + 2 ] );
+		maxX = Math.max( maxX, f[ o + A ], f[ o + B ], f[ o + C ] );
+		maxY = Math.max( maxY, f[ o + A + 1 ], f[ o + B + 1 ], f[ o + C + 1 ] );
+		maxZ = Math.max( maxZ, f[ o + A + 2 ], f[ o + B + 2 ], f[ o + C + 2 ] );
+
+	}
+
+	out[ off ] = minX; out[ off + 1 ] = minY; out[ off + 2 ] = minZ;
+	out[ off + 3 ] = maxX; out[ off + 4 ] = maxY; out[ off + 5 ] = maxZ;
+
+}
+
+const folded = new Float32Array( 12 );
+
+/**
+ * One node of a bottom-up refit: its own bounds into `bounds[ ( i - base ) * 6 ]`, and an inner
+ * node's two child boxes rewritten from theirs. A leaf folded into its parent takes its box from
+ * its triangles.
+ */
+function refitNode( nodes, i, acc, bounds, base ) {
+
+	const bvhF = nodeF( nodes, i );
+	const idx = nodeIdx( nodes, i );
+	const o = nodeBase( nodes, i );
+	const b = ( i - base ) * 6;
+	const marker = idx[ o + 3 ];
+
+	if ( marker === LEAF_MARKER ) {
+
+		triangleBounds( acc, idx[ o ], idx[ o + 1 ], bounds, b );
+		return;
+
+	}
+
+	if ( marker === BLAS_POINTER_MARKER ) {
+
+		// BLAS-pointer leaf (TLAS): the BLAS root's bounds are in the instance's own
+		// space, and the TLAS sorts on world bounds — so carry the box out through the
+		// leaf's transform. Slots 4..15 are the rows of world-to-object; the eight
+		// corners go back the other way through its inverse.
+		transformBoundsToWorld( bvhF, o, bounds, ( idx[ o ] - base ) * 6, bounds, b );
+		return;
+
+	}
+
+	const right = idx[ o + 7 ];
+	let ls = bounds, lb = ( marker - base ) * 6;
+	if ( isFoldedRef( marker ) ) {
+
+		triangleBounds( acc, foldedFirst( marker ), foldedCount( marker ), folded, 0 );
+		ls = folded; lb = 0;
+
+	}
+
+	let rs = bounds, rb = ( right - base ) * 6;
+	if ( isFoldedRef( right ) ) {
+
+		triangleBounds( acc, foldedFirst( right ), foldedCount( right ), folded, 6 );
+		rs = folded; rb = 6;
+
+	}
+
+	bvhF[ o ] = ls[ lb ]; bvhF[ o + 1 ] = ls[ lb + 1 ]; bvhF[ o + 2 ] = ls[ lb + 2 ];
+	bvhF[ o + 4 ] = ls[ lb + 3 ]; bvhF[ o + 5 ] = ls[ lb + 4 ]; bvhF[ o + 6 ] = ls[ lb + 5 ];
+	bvhF[ o + 8 ] = rs[ rb ]; bvhF[ o + 9 ] = rs[ rb + 1 ]; bvhF[ o + 10 ] = rs[ rb + 2 ];
+	bvhF[ o + 12 ] = rs[ rb + 3 ]; bvhF[ o + 13 ] = rs[ rb + 4 ]; bvhF[ o + 14 ] = rs[ rb + 5 ];
+
+	for ( let a = 0; a < 3; a ++ ) {
+
+		bounds[ b + a ] = Math.min( ls[ lb + a ], rs[ rb + a ] );
+		bounds[ b + 3 + a ] = Math.max( ls[ lb + 3 + a ], rs[ rb + 3 + a ] );
+
+	}
+
+}
+
 export class BVHRefitter {
 
 	constructor() {
@@ -254,102 +349,9 @@ export class BVHRefitter {
 
 		}
 
-		const bounds = this._bounds;
-		const endNode = startNode + nodeCount;
 		const nodes = nodeAccess( bvhData );
-
-		for ( let i = endNode - 1; i >= startNode; i -- ) {
-
-			const bvhF = nodeF( nodes, i );
-			const idx = nodeIdx( nodes, i );
-			const o = nodeBase( nodes, i );
-			const b = ( i - startNode ) * 6; // bounds indexed relative to BLAS start
-
-			if ( idx[ o + 3 ] === LEAF_MARKER ) {
-
-				const triOffset = idx[ o ];
-				const triCount = idx[ o + 1 ];
-
-				let minX = Infinity, minY = Infinity, minZ = Infinity;
-				let maxX = - Infinity, maxY = - Infinity, maxZ = - Infinity;
-
-				for ( let t = 0; t < triCount; t ++ ) {
-
-					const gi = triOffset + t;
-					const triFloats = acc.chunked ? acc.f.chunkFor( gi ) : acc.f;
-					const tOff = acc.chunked ? acc.chunked.baseOf( gi ) : gi * FPT;
-					const ax = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_A_OFFSET ];
-					const ay = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_A_OFFSET + 1 ];
-					const az = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_A_OFFSET + 2 ];
-					const bx = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_B_OFFSET ];
-					const by = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_B_OFFSET + 1 ];
-					const bz = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_B_OFFSET + 2 ];
-					const cx = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_C_OFFSET ];
-					const cy = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_C_OFFSET + 1 ];
-					const cz = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_C_OFFSET + 2 ];
-
-					minX = Math.min( minX, ax, bx, cx );
-					minY = Math.min( minY, ay, by, cy );
-					minZ = Math.min( minZ, az, bz, cz );
-					maxX = Math.max( maxX, ax, bx, cx );
-					maxY = Math.max( maxY, ay, by, cy );
-					maxZ = Math.max( maxZ, az, bz, cz );
-
-				}
-
-				bounds[ b ] = minX;
-				bounds[ b + 1 ] = minY;
-				bounds[ b + 2 ] = minZ;
-				bounds[ b + 3 ] = maxX;
-				bounds[ b + 4 ] = maxY;
-				bounds[ b + 5 ] = maxZ;
-
-			} else {
-
-				// Inner node — child indices are absolute, but bounds index relative to startNode
-				const leftIdx = idx[ o + 3 ];
-				const rightIdx = idx[ o + 7 ];
-				const lb = ( leftIdx - startNode ) * 6;
-				const rb = ( rightIdx - startNode ) * 6;
-
-				const lMinX = bounds[ lb ];
-				const lMinY = bounds[ lb + 1 ];
-				const lMinZ = bounds[ lb + 2 ];
-				const lMaxX = bounds[ lb + 3 ];
-				const lMaxY = bounds[ lb + 4 ];
-				const lMaxZ = bounds[ lb + 5 ];
-
-				const rMinX = bounds[ rb ];
-				const rMinY = bounds[ rb + 1 ];
-				const rMinZ = bounds[ rb + 2 ];
-				const rMaxX = bounds[ rb + 3 ];
-				const rMaxY = bounds[ rb + 4 ];
-				const rMaxZ = bounds[ rb + 5 ];
-
-				bvhF[ o ] = lMinX;
-				bvhF[ o + 1 ] = lMinY;
-				bvhF[ o + 2 ] = lMinZ;
-				bvhF[ o + 4 ] = lMaxX;
-				bvhF[ o + 5 ] = lMaxY;
-				bvhF[ o + 6 ] = lMaxZ;
-
-				bvhF[ o + 8 ] = rMinX;
-				bvhF[ o + 9 ] = rMinY;
-				bvhF[ o + 10 ] = rMinZ;
-				bvhF[ o + 12 ] = rMaxX;
-				bvhF[ o + 13 ] = rMaxY;
-				bvhF[ o + 14 ] = rMaxZ;
-
-				bounds[ b ] = Math.min( lMinX, rMinX );
-				bounds[ b + 1 ] = Math.min( lMinY, rMinY );
-				bounds[ b + 2 ] = Math.min( lMinZ, rMinZ );
-				bounds[ b + 3 ] = Math.max( lMaxX, rMaxX );
-				bounds[ b + 4 ] = Math.max( lMaxY, rMaxY );
-				bounds[ b + 5 ] = Math.max( lMaxZ, rMaxZ );
-
-			}
-
-		}
+		// Bounds indexed relative to the BLAS start; child indices are absolute.
+		for ( let i = startNode + nodeCount - 1; i >= startNode; i -- ) refitNode( nodes, i, acc, this._bounds, startNode );
 
 	}
 
@@ -374,126 +376,8 @@ export class BVHRefitter {
 
 		}
 
-		const bounds = this._bounds;
 		const nodes = nodeAccess( bvhData );
-
-		// Reverse iteration: bottom-up in pre-order layout
-		for ( let i = nodeCount - 1; i >= 0; i -- ) {
-
-			const bvhF = nodeF( nodes, i );
-			const idx = nodeIdx( nodes, i );
-			const o = nodeBase( nodes, i );
-			const b = i * 6;
-
-			const marker = idx[ o + 3 ];
-
-			if ( marker === LEAF_MARKER ) {
-
-				// Triangle leaf: compute AABB from triangles
-				const triOffset = idx[ o ];
-				const triCount = idx[ o + 1 ];
-
-				let minX = Infinity, minY = Infinity, minZ = Infinity;
-				let maxX = - Infinity, maxY = - Infinity, maxZ = - Infinity;
-
-				for ( let t = 0; t < triCount; t ++ ) {
-
-					const gi = triOffset + t;
-					const triFloats = acc.chunked ? acc.f.chunkFor( gi ) : acc.f;
-					const tOff = acc.chunked ? acc.chunked.baseOf( gi ) : gi * FPT;
-
-					// Position A
-					const ax = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_A_OFFSET ];
-					const ay = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_A_OFFSET + 1 ];
-					const az = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_A_OFFSET + 2 ];
-					// Position B
-					const bx = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_B_OFFSET ];
-					const by = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_B_OFFSET + 1 ];
-					const bz = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_B_OFFSET + 2 ];
-					// Position C
-					const cx = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_C_OFFSET ];
-					const cy = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_C_OFFSET + 1 ];
-					const cz = triFloats[ tOff + TRIANGLE_DATA_LAYOUT.POSITION_C_OFFSET + 2 ];
-
-					minX = Math.min( minX, ax, bx, cx );
-					minY = Math.min( minY, ay, by, cy );
-					minZ = Math.min( minZ, az, bz, cz );
-					maxX = Math.max( maxX, ax, bx, cx );
-					maxY = Math.max( maxY, ay, by, cy );
-					maxZ = Math.max( maxZ, az, bz, cz );
-
-				}
-
-				bounds[ b ] = minX;
-				bounds[ b + 1 ] = minY;
-				bounds[ b + 2 ] = minZ;
-				bounds[ b + 3 ] = maxX;
-				bounds[ b + 4 ] = maxY;
-				bounds[ b + 5 ] = maxZ;
-
-			} else if ( marker === BLAS_POINTER_MARKER ) {
-
-				// BLAS-pointer leaf (TLAS): the BLAS root's bounds are in the instance's own
-				// space, and the TLAS sorts on world bounds — so carry the box out through the
-				// leaf's transform. Slots 4..15 are the rows of world-to-object; the eight
-				// corners go back the other way through its inverse.
-				const blasRoot = idx[ o ];
-				const br = blasRoot * 6;
-				transformBoundsToWorld( bvhF, o, bounds, br, bounds, b );
-
-			} else {
-
-				// Inner node: union children bounds (already computed since we iterate in reverse)
-				const leftIdx = idx[ o + 3 ];
-				const rightIdx = idx[ o + 7 ];
-				const lb = leftIdx * 6;
-				const rb = rightIdx * 6;
-
-				const lMinX = bounds[ lb ];
-				const lMinY = bounds[ lb + 1 ];
-				const lMinZ = bounds[ lb + 2 ];
-				const lMaxX = bounds[ lb + 3 ];
-				const lMaxY = bounds[ lb + 4 ];
-				const lMaxZ = bounds[ lb + 5 ];
-
-				const rMinX = bounds[ rb ];
-				const rMinY = bounds[ rb + 1 ];
-				const rMinZ = bounds[ rb + 2 ];
-				const rMaxX = bounds[ rb + 3 ];
-				const rMaxY = bounds[ rb + 4 ];
-				const rMaxZ = bounds[ rb + 5 ];
-
-				// Write left child AABB into bvhData
-				bvhF[ o ] = lMinX;
-				bvhF[ o + 1 ] = lMinY;
-				bvhF[ o + 2 ] = lMinZ;
-				// o+3 = leftChildIdx (preserved)
-				bvhF[ o + 4 ] = lMaxX;
-				bvhF[ o + 5 ] = lMaxY;
-				bvhF[ o + 6 ] = lMaxZ;
-				// o+7 = rightChildIdx (preserved)
-
-				// Write right child AABB into bvhData
-				bvhF[ o + 8 ] = rMinX;
-				bvhF[ o + 9 ] = rMinY;
-				bvhF[ o + 10 ] = rMinZ;
-				// o+11 = 0 padding
-				bvhF[ o + 12 ] = rMaxX;
-				bvhF[ o + 13 ] = rMaxY;
-				bvhF[ o + 14 ] = rMaxZ;
-				// o+15 = 0 padding
-
-				// Store this node's bounds as union of children
-				bounds[ b ] = Math.min( lMinX, rMinX );
-				bounds[ b + 1 ] = Math.min( lMinY, rMinY );
-				bounds[ b + 2 ] = Math.min( lMinZ, rMinZ );
-				bounds[ b + 3 ] = Math.max( lMaxX, rMaxX );
-				bounds[ b + 4 ] = Math.max( lMaxY, rMaxY );
-				bounds[ b + 5 ] = Math.max( lMaxZ, rMaxZ );
-
-			}
-
-		}
+		for ( let i = nodeCount - 1; i >= 0; i -- ) refitNode( nodes, i, acc, this._bounds, 0 );
 
 	}
 

package/src/Processor/ChunkedRecords.js

@@ -134,7 +134,7 @@ export class ChunkedRecords {
 		const lanes = Math.min( this.recordsPerChunk, this.recordCount - from ) * this.lanesPerRecord;
 		const chunk = this.chunks[ k ] = allocChunk( this.LaneType, lanes, this.shared );
 
-		for ( const v of this._views ) {
+		for ( const v of this._liveViews() ) {
 
 			v.chunks[ k ] = new v.LaneType( chunk.buffer, chunk.byteOffset, chunk.byteLength / v.LaneType.BYTES_PER_ELEMENT );
 
@@ -147,7 +147,7 @@ export class ChunkedRecords {
 	/** Allocate whatever a lazy fill never touched, so the result behaves like an eager one. */
 	materializeAll() {
 
-		for ( let k = 0; k < this.chunks.length; k ++ ) if ( ! this.chunks[ k ] ) this._materialize( k );
+		for ( let k = 0; k < this.chunks.length; k ++ ) if ( ! this.chunks[ k ] && ! this._spilled?.has( k ) ) this._materialize( k );
 		return this;
 
 	}
@@ -177,7 +177,104 @@ export class ChunkedRecords {
 	chunkFor( record ) {
 
 		const k = ( record / this.recordsPerChunk ) | 0;
-		return this.chunks[ k ] ?? this._materialize( k );
+		const chunk = this.chunks[ k ];
+		if ( chunk ) return chunk;
+		if ( this._spilled?.has( k ) ) throw new Error( `ChunkedRecords: chunk ${k} is on disk — await ensureResident() before touching record ${record}` );
+		return this._materialize( k );
+
+	}
+
+	/**
+	 * Writes every resident chunk to `store` and lets go of it; {@link ensureResident} reads it
+	 * back. A spilled chunk's accessors throw rather than read zeros. Every other holder of a chunk
+	 * (a worker, a view) keeps it alive, so spill only what nothing else holds.
+	 * @param {{write: function(number, ArrayBufferView): Promise<void>, read: function(number, number): Promise<ArrayBuffer>}} store
+	 * @param {{keep?: function(number): boolean}} [options] - chunks that stay in memory
+	 * @returns {Promise<number>} bytes let go
+	 */
+	async spill( store, { keep = () => false } = {} ) {
+
+		if ( this._owner ) throw new Error( 'ChunkedRecords: spill the storage, not a view of it' );
+		this._store = store;
+		this._spilled ??= new Set();
+		let bytes = 0;
+
+		for ( let k = 0; k < this.chunks.length; k ++ ) {
+
+			const chunk = this.chunks[ k ];
+			if ( ! chunk || keep( k ) ) continue;
+			await store.write( k, chunk );
+			this.chunks[ k ] = null;
+			for ( const v of this._liveViews() ) v.chunks[ k ] = null;
+			this._spilled.add( k );
+			bytes += chunk.byteLength;
+
+		}
+
+		return bytes;
+
+	}
+
+	/** Writes one chunk to `store` and lets go of it — a build hands chunks over as it finishes them. */
+	async spillChunk( k, store ) {
+
+		const chunk = this.chunks[ k ];
+		if ( ! chunk ) return 0;
+		this._store = store;
+		this._spilled ??= new Set();
+		await store.write( k, chunk );
+		this.chunks[ k ] = null;
+		for ( const v of this._liveViews() ) v.chunks[ k ] = null;
+		this._spilled.add( k );
+		return chunk.byteLength;
+
+	}
+
+	/** Reads the spilled chunks under records [start, start+count) back into memory. */
+	async ensureResident( start = 0, count = this.recordCount ) {
+
+		if ( ! this._spilled?.size || count <= 0 ) return;
+		const first = ( start / this.recordsPerChunk ) | 0;
+		const last = ( ( start + count - 1 ) / this.recordsPerChunk ) | 0;
+
+		for ( let k = first; k <= last; k ++ ) {
+
+			if ( ! this._spilled.has( k ) ) continue;
+			const lanes = Math.min( this.recordsPerChunk, this.recordCount - k * this.recordsPerChunk ) * this.lanesPerRecord;
+			const bytes = await this._store.read( k, lanes * this.LaneType.BYTES_PER_ELEMENT );
+			const chunk = allocChunk( this.LaneType, lanes, this.shared );
+			new Uint8Array( chunk.buffer, chunk.byteOffset, chunk.byteLength ).set( new Uint8Array( bytes ) );
+			this.chunks[ k ] = chunk;
+			for ( const v of this._liveViews() ) v.chunks[ k ] = new v.LaneType( chunk.buffer, chunk.byteOffset, chunk.byteLength / v.LaneType.BYTES_PER_ELEMENT );
+			this._spilled.delete( k );
+
+		}
+
+	}
+
+	/** @private */
+	_liveViews() {
+
+		if ( this._weakViews ) this._weakViews = this._weakViews.filter( ref => ref.deref() );
+		return [ ...( this._views ?? [] ), ...( this._weakViews ?? [] ).map( ref => ref.deref() ) ];
+
+	}
+
+	/** Whether records [start, start+count) are all in memory. */
+	isResident( start = 0, count = this.recordCount ) {
+
+		if ( ! this._spilled?.size || count <= 0 ) return true;
+		const first = ( start / this.recordsPerChunk ) | 0;
+		const last = ( ( start + count - 1 ) / this.recordsPerChunk ) | 0;
+		for ( let k = first; k <= last; k ++ ) if ( this._spilled.has( k ) ) return false;
+		return true;
+
+	}
+
+	/** Chunks currently on disk. */
+	get spilledChunks() {
+
+		return this._spilled?.size ?? 0;
 
 	}
 
@@ -278,6 +375,87 @@ export class ChunkedRecords {
 
 	}
 
+	/**
+	 * A lazy store taking more records without moving the ones it has: new chunks are added and
+	 * only a short last chunk is reallocated. A build streaming chunks to disk cannot copy the
+	 * store, since earlier chunks may already be gone.
+	 */
+	grow( recordCount ) {
+
+		if ( recordCount <= this.recordCount ) return;
+		const last = this.chunks.length - 1;
+		const lanesFor = ( k, count ) => Math.min( this.recordsPerChunk, count - k * this.recordsPerChunk ) * this.lanesPerRecord;
+		const old = this.chunks[ last ];
+
+		this.recordCount = recordCount;
+		for ( const v of this._liveViews() ) v.recordCount = recordCount;
+		if ( this.chunks.length * this.recordsPerChunk < recordCount ) {
+
+			this.chunks.length = Math.ceil( recordCount / this.recordsPerChunk );
+			for ( const v of this._liveViews() ) v.chunks.length = this.chunks.length;
+
+		}
+
+		if ( old && old.length < lanesFor( last, recordCount ) ) {
+
+			const chunk = allocChunk( this.LaneType, lanesFor( last, recordCount ), this.shared );
+			chunk.set( old );
+			this.chunks[ last ] = chunk;
+			for ( const v of this._liveViews() ) v.chunks[ last ] = new v.LaneType( chunk.buffer, chunk.byteOffset, chunk.byteLength / v.LaneType.BYTES_PER_ELEMENT );
+
+		} else if ( ! old && this._spilled?.has( last ) ) {
+
+			throw new Error( `ChunkedRecords: cannot grow past chunk ${last}, which is on disk` );
+
+		}
+
+	}
+
+	/**
+	 * Lets go of the chunks lying wholly inside records [start, start+count) without writing them
+	 * anywhere: their contents are about to be rewritten, and the next access allocates them again.
+	 * @returns {number} bytes let go
+	 */
+	release( start, count ) {
+
+		const rpc = this.recordsPerChunk;
+		const first = Math.ceil( start / rpc );
+		const end = Math.floor( ( start + count ) / rpc );
+		let bytes = 0;
+		for ( let k = first; k < end; k ++ ) {
+
+			const chunk = this.chunks[ k ];
+			if ( ! chunk ) continue;
+			bytes += chunk.byteLength;
+			this.chunks[ k ] = undefined;
+			for ( const v of this._liveViews() ) v.chunks[ k ] = undefined;
+
+		}
+
+		return bytes;
+
+	}
+
+	/** {@link trimTo} on this store itself, keeping what it has spilled. */
+	trimInPlace( recordCount ) {
+
+		const needed = Math.max( 1, Math.ceil( recordCount / this.recordsPerChunk ) );
+		const lanes = Math.max( 0, ( recordCount - ( needed - 1 ) * this.recordsPerChunk ) * this.lanesPerRecord );
+		const stores = [ this, ...this._liveViews() ];
+		for ( const s of stores ) {
+
+			s.chunks.length = needed;
+			const last = s.chunks[ needed - 1 ];
+			if ( last && last.length > lanes ) s.chunks[ needed - 1 ] = last.subarray( 0, lanes );
+			s.recordCount = recordCount;
+
+		}
+
+		if ( this._spilled ) for ( const k of this._spilled ) if ( k >= needed ) this._spilled.delete( k );
+		return this;
+
+	}
+
 	/** A parallel set of views of another lane type over the same memory (u32 records read as f32). */
 	viewAs( LaneType ) {
 
@@ -287,14 +465,21 @@ export class ChunkedRecords {
 			this.recordCount, this.lanesPerRecord, this.recordsPerChunk
 		);
 		v.LaneType = LaneType;
+		// A view never allocates for itself: a chunk it finds missing is the storage's to create.
+		v._owner = this;
 
 		if ( this._views ) {
 
-			v._owner = this;
 			this._views.push( v );
 
+		} else {
+
+			// Weakly: callers take short-lived views, and a spill must still reach the lasting ones.
+			( this._weakViews ??= [] ).push( new WeakRef( v ) );
+
 		}
 
+		v._spilled = this._spilled ??= new Set();
 		return v;
 
 	}

package/src/Processor/EmissiveTriangleBuilder.js

@@ -90,6 +90,14 @@ export class EmissiveTriangleBuilder {
 
 		for ( let i = 0; i < triangleCount; i ++ ) {
 
+			// A chunk a build spilled holds no emitters: that is why it could be spilled.
+			if ( chunked?.spilledChunks && ! chunked.isResident( i, 1 ) ) {
+
+				i = ( ( ( i / chunked.recordsPerChunk ) | 0 ) + 1 ) * chunked.recordsPerChunk - 1;
+				continue;
+
+			}
+
 			const triU = chunked ? chunked.chunkFor( i ) : flatU;
 			const triFloats = chunked ? chunkedF.chunkFor( i ) : flatF;
 			const baseOffset = chunked ? chunked.baseOf( i ) : i * FLOATS_PER_TRIANGLE;